37 innovations from Bar-Ilan University, available for licensing, co-investment, or spin-out through BIRAD.
Zalevsky Zeev
Intraocular pressure (IOP) measurements comprise an essential tool in modern medicine for the early diagnosis of glaucoma, the second leading cause of human blindness. The world's highest prevalence of glaucoma is in low-income countries. Current diagnostic methods require experience in running expensive equipment as well as the use of anesthetic eye drops. We present herein a remote photonic IOP biomonitoring method based on deep learning of secondary speckle patterns, captured by a fast camera, that are reflected from eye sclera stimulated by an external sound wave. By combining speckle pattern analysis with deep learning, high precision measurements are possible. The method was tested under artificially varying eye pressures on a series of 24 pig eyeballs, found to be similar to human eyes. As a low-cost procedure, it has the potential to meet clinical needs in low- and middle-income countries and at points of care everywhere.
Singer Gonen
Transformers play a central role in modern artificial intelligence, yet their susceptibility to adversarial perturbations raises serious reliability concerns. Current defenses, such as adversarial training, are often computationally expensive and tailored to specific attack vectors, while formal verification methods remain difficult to scale. In this work, we propose RAHP (Robustness-Aware Head Pruning), a framework that enhances the intrinsic robustness of Transformers by selectively removing attention heads that contribute to model adversarial vulnerability. Unlike standard pruning, which often degrades robustness, RAHP guides the pruning process using a composite score of two complementary signals: (i) Fisher Information, which preserves task accuracy, and (ii) CLEVER, a sensitivity-based proxy derived from the local Lipschitz constant that estimates the model’s vulnerability to perturbations. By pruning attention heads that exhibit high adversarial sensitivity, RAHP steers the model toward a more stable decision boundary without the need for costly adversarial retraining. Extensive experiments demonstrate that RAHP yields compact models that are not only efficient but also more resistant to a wide variety of attacks compared to standard pruning and regularization baselines. These results suggest that incorporating local stability criteria into the pruning process provides a scalable and attack-agnostic pathway toward robust and efficient Transformer models.
Adi Makmal
Graph analysis constitutes a foundational framework across modern data science, infrastructure engineering, and computational modeling, serving as the core mathematical architecture for mapping complex relational dependencies in real-world systems The present invention introduces a novel, scalable quantum computational framework and algorithmic process designed to generate exponentially large, non-trivial mathematical graphs that are directly mapped into highly compact quantum physical operators. This is in contrast to conventional techniques that yield complex, unmanageable operator expressions requiring an exponential number of terms relative to the number of system qubits. The process operates by first constructing a highly symmetric, spectrally-solvable structural backbone known as a 'Skeleton Laplacian Hamiltonian' utilizing a restricted Pauli operator subset (consisting purely of tensor products of Identity and Pauli-X operators, excluding the all-identity string). This structural backbone maps an unweighted d-regular skeleton graph using an extremely sparse allocation of only (d + 1) Pauli string terms, thereby completely decoupling the physical description length from the overall network size. To bypass structural and spectral triviality while retaining strict logarithmic scaling, the invention establishes a localized embedding process. Small graphs are encoded through a standard basis element outer-product structure and directly injected as localized structural modifications into specified coordinates of the global Skeleton backbone. This mechanism disrupts the macro-level structural symmetries of the skeleton network in a controlled manner, successfully creating complex and highly scalable graph architectures. Consequently, the complete global graph matrix Laplacian operator is accurately expressed on real physical quantum hardware utilizing a strictly constrained allocation of only q = log(n) qubits and an efficient polylogarithmic O(polylog(n)) number of physical Pauli string operators. This method unlocks the ability to analyze exponentially large graphs on quantum hardware of exponentially large graphs using quantum algorithms, extending the reach of quantum algorithms into scale where classical tools are no longer applicable, with applications in optimization processes and large scale data analysis.
Yavits Leonid
We present SquiBNet, a deep learning framework for large-scale bacterial pathogen identification directly from raw Oxford Nanopore Technologies (ONT) electrical signals, termed squiggles. Conventional nanopore pipelines require computationally expensive basecalling and reference alignment before classification. These stages introduce latency and preclude real-time operation on resource-constrained edge platforms. SquiBNet bypasses both stages and classifies organisms directly in signal space. The architecture combines a one-dimensional ConvNeXt-Base backbone with four Transformer encoder blocks. The convolutional backbone captures local signal structure. The Transformer head models long-range dependencies across the squiggle. Across 256 clinically relevant bacterial species, SquiBNet attains 82.07% test accuracy. This represents a 51.5 percentage-point absolute gain, approximately 2.7 times, over the strongest prior-art baseline adapted to the same task. Training relies on a validated large-scale simulated dataset constructed with Badread, Squiggulator, and slow5tools, comprising 18.5 million signal windows. Simulated signals were validated against real nanopore data through signal-to-noise ratio analysis and visual comparison. On the NVIDIA Jetson Orin NX edge platform, SquiBNet achieves a single-sample latency of 41.5 ms and sustains 32.5 samples per second at batch size 32, with negligible preprocessing overhead. The framework supports concurrent real-time classification of multiple nanopore channels, enabling point-of-care deployment within Read Until adaptive sequencing.
Cohen Cyrille
The invention involves the computational design and development of novel constant regions for human T-cell receptors (TCRs), termed "Structurally Enhanced TCR" (SET). By introducing a set of strategic mutations into the TCR constant domains, the SET design improves receptor stability, enhances surface expression, increases functional avidity, and ensures preferential pairing of the α and β chains, minimizing mispairing with endogenous TCRs. This innovation offers a universal platform to optimize T-cell therapies for cancer and infectious diseases without requiring additional gene editing.
Cohen Haim
To explore the role of protein post-translational modifications on lifespan and healthspan , we developed the PHARAOH computational tool based on the 100-fold differences in longevity within the mammalian class. Analyzing acetylome/phosphorylome and proteome data across 107 mammalian species identified 482 and 695 significant longevity-associated acetylated lysine residues in mice and humans, respectively. In addition, we have recognized 2115 longevity associated p phosphorylations. In regards to acetylations, these sites include acetylated lysines in short-lived mammals that were replaced by permanent acetylation or deacetylation mimickers, glutamine or arginine, respectively, in long-lived mammals. Conversely, glutamine or arginine residues in short-lived mammals were replaced by reversibly acetylated lysine in long-lived mammals. For phosphorylations, site these sites include phosphorylated Serine (S), Tyrosine (Y) and threonine (T) or their replacement to aspartic acid (D) or glutamic acid (E)or Alanine (A) and Y to phenylalanine (Y to F). Pathway analyses of the acetylation sites highlighted the involvement of mitochondrial translation, cell cycle, fatty acid oxidation, transsulfuration, DNA repair and others in longevity. A validation assay showed that substitution of lysine 386 with arginine in mouse cystathionine beta synthase, to attain the human sequence, increased the pro-longevity activity of this enzyme. Likewise, replacing the human ubiquitin-specific peptidase 10 acetylated lysine 714 with arginine as in short-lived mammals, reduced its anti-neoplastic function. These findings provide a computational tool for identifying modifications that control longer healthy life and potential interventions to extend human healthspan.
Ozana, Nisan
The present invention relates to a non-invasive optical system and method for determining intra-facial connectivity and face–brain connectivity using full-field speckle analysis. The system is configured to illuminate substantially the entire facial surface using coherent or partially coherent light and to capture dynamic speckle patterns generated by facial tissues, including microvascular activity, neuromuscular activity, and micromovements. The captured speckle data are processed to extract temporal and spatial features including, but not limited to, speckle contrast, intensity fluctuations, temporal decorrelation rates, autocorrelation functions, frequency-domain characteristics, phase relationships, and vibration-related parameters. The facial surface is segmented into multiple regions of interest, each generating time-resolved speckle-derived signals. Functional connectivity within the face (intra-facial connectivity) is computed by calculating correlation, coherence, mutual information, phase synchronization, or other statistical or signal-processing metrics between time-series signals derived from different facial regions. These computations generate connectivity matrices and network representations reflecting dynamic coupling patterns across the face. In certain embodiments, brain activity signals are acquired simultaneously or sequentially using electrical, optical, acoustic, or multimodal sensing techniques. The facial speckle-derived signals are synchronized with brain activity measurements to calculate face–brain connectivity using cross-correlation, coherence analysis, causality estimation, or other cross-modal coupling methods. The resulting connectivity maps may reflect functional interactions between facial dynamics and neural activity. The system may further include data processing modules, machine learning algorithms, and classification models configured to identify patterns associated with neurological, psychiatric, or neuromuscular conditions. The invention may be implemented as a stationary, portable, or wearable device and may operate in real time or offline analysis modes. By providing quantitative measurements of intra-facial and face–brain connectivity, the invention enables objective assessment, monitoring, and biomarker development for applications including mental health disorders, addiction, depression, cognitive impairment, neurodegenerative diseases, and facial paralysis of different etiologies.
Yavits Leonid
A dedicated hardware unit embedded within the shared substrate of a heterogeneous multi-chiplet semiconductor package. It continuously monitors the package's physical health using thermal, mechanical strain, and electrical aging sensors, runs embedded AI inference to predict failures, and autonomously adjusts power and frequency to prevent damage and optimize the operation — all from within the package substrate itself.
Keshet Joseph
There is provided a system for computing a secure statistical classifier, comprising: at least one hardware processor executing a code for: accessing code instructions of an untrained statistical classifier, accessing a training dataset, accessing a plurality of cryptographic keys, creating a plurality of instances of the untrained statistical classifier, creating a plurality of trained sub-classifiers by training each of the plurality of instances of the untrained statistical classifier by iteratively adjusting adjustable classification parameters of the respective instance of the untrained statistical classifier according to a portion of the training data serving as input and a corresponding ground truth label, and at least one unique cryptographic key of the plurality of cryptographic keys, wherein the adjustable classification parameters of each trained sub-classifier have unique values computed according to corresponding at least one unique cryptographic key, and providing the statistical classifier, wherein the statistical classifier includes the plurality of trained sub-classifiers.
Levy Oren
We introduce a unique and novel customizable 3D interface for producing scalable, biomimetic artificial reefs (ARs), utilizing real data collected from coral ecosystems. This interface employs 3D technologies, 3D imaging with AI generative models, and 3D printing, to extract core reef characteristics, which can be translated and digitized into a 3D printed reef. The advantages of 3D printing lie in providing customized tools by which to integrate the vital details of natural reefs, such as rugosity and complexity, into a sustainable manufacturing process. This methodology can offer economic solutions for developing both small and large-scale biomimetic structures for a variety of restoration situations, that closely resemble the coral reefs they intend to support. Our method consists of the following steps: 1. 3D photogrammetric scan. 2. Generating 3D models based on the scans and prior knowledge from reefs. 3. Printing the generated 3D models. Artificial reefs are designed to resemble natural reefs to the highest degree possible, maximizing restoration efforts both ecologically and aesthetically. Coral reefs are mapped in 3D by diver-based or platform passed using AUV photogrammetry. This technology enables the production of highly detailed 3D models of the substrate and detection of the sessile organisms that inhabit the reef. Conducting photogrammetric surveys in areas which are designated for reef reformation with our ARs, is highly beneficial, as it identifies which natural reef structures harbor the large biodiversity as well as depicting their numerical composition (diversity) within the reef structure using unique AI. CAD design platforms (i.e., Rhinoceros© and Grasshopper©) create biomimetic or bio-inspired designs, using ceramic 3D printing (3DP). Incorporating our 3D model (images) provides a natural foundation to interactively customize it to fit the needs of any type of reef geographically, depth, etc. or refine the biomimetic design. The 3D models are further analyzed geometrically using advanced data-analysis tools, to extract the general features and characteristics that will lead to a successful AR. We offer plug-ins for designing artificial structures that consolidate algorithms based on the formation of a coral reef structure and our eDNA information that can predict the types of biodiversity it may maintain/accommodate. An eDNA and metabarcoding package is combined to monitor and extract key biological and ecological information about coral reefs, indicating its pivotal potential as an evaluation tool for ARs and reef restoration success. Removable appendages incorporated in the desing of our ARs, will be used for eDNA biomass (organism) surveys, alongside seawater samples, without interfering with the restoration process. As the information extracted from eDNA is broadly expansive, it can be utilized to predict biodiversity outcomes of the 3D printed AR based on the 3D imaged reef. Furthermore, eDNA data can help to understand what characteristics of the 3D modeled reef are related to the diversity of organisms that inhabit it, which guide further the fabrication of the AR. eDNA is a useful tool to observe these hard to identify communities and to understand which species benefit most from the AR structure. This information will be collected to understand the community composition, abundance, and richness, available through eDNA analysis. Data collected from coral reefs, using eDNA and 3D imaging, can reinforce their resilience through establishing baselines, monitoring, and evaluation of restoration activities. Combining these data-acquisition tools with 3DP offers a holistic approach to manufacturing biomimetic ARs that are tailor-made to any coral reef worldwide. Eventually leading to an entirely data-driven interface utilizing parametric design software and machine-learning tools to curate an algorithm for customizing ARs, according to the specific, desired characteristics of the reef, such as reef structure/type, biodiversity, depth, coral morphology, etc. Moreover, the algorithm will automate the optimization of AR designs according to the data it is supplied from eDNA, 3D imaging, and other monitoring surveys. This methodology would make it possible to determine the precise design parameters needed to construct an AR, provide a baseline for the expected biodiversity that could accumulate on 3D printed ARs, and ensure no excess waste in the manufacturing process. The development of sustainable large-scale and long-term projects that can provide key social and economic benefits will be a demand of the future. When marine restoration projects manage to meet these requirements, they are able to achieve restoration goals together with social and economic change. Novelty Point out the novel aspects, of your invention (what is new about it, or what are its new features) in detail. Please emphasize the non-obvious/unpredicted aspects of the invention. The novelty of this invention lies in the fabric of the following ingredients: 1. Using unique and novel AI algorithm to map the local biodiversity to select “hotspot” biomimicry reefs. 2. Using novel generative AI methods to generate 3D models based on 3D photogrammetric scans of the environment. 3. Using a unique and novel cross-examination of the AI analyzing with eDNA – predictive biodiversity of 3D printed reef 4. Using unique and novel translation of 3D photogrammetric scanning of the reef into a 3D printable shape. 5. Using unique and novel algorithm to translate the 3D shape into a movement of the 3D printing of pasty materials, such as clay, aligned with calcium carbonate 4f. Advantages of the Invention Describe the advantages of your invention over the conventional manner for solving the problem, describing how and why your invention does it better: The advantages of the invention are: 1. Biomimicry/natural reef replication 2. High rate of biological success 3. Eco-materials using clay 4. Tailor made for any location. 5. Cost effective with digital manufacturing 6. Integration of a process by using multi-disciplinary approach: 7. Incorporation of reef and environmental characteristics 8. Large-scale solutions using advanced scaling and fabrication
Glickman, Oren
We offer a low-cost, relatively small-size device based on imaging and AI technology that enables assessing key quality indicators of tomato fruits (fruit size, weight, and firmness, as well as total soluble solides (TSS), pH, acidity, Vitamin-C, and lycopene contents in the fruit) without having to collect the fruits and take them to the lab for analyses or even touch the fruits. Moreover, the device can provide quality indicators for many fruits simultaneously from a single shot in a speedy process. The latter feature enables the assessment of quality indicators in many tomato fruits in a very short time. It can be used by (1) various food industry companies that want to use a cheap and speedy process to classify their post-harvest fresh produce per their quality or select the most suitable fruits for their specific purposes (products). The device can also be used by (2) farmers to improve the quality of their fruits by tracking changes in quality indicators along the ripening stages and taking specific actions to influence such quality changes or by just integrating the device into autonomous systems (robots that are being implemented in the agricultural sector these days to make the harvest process an automatic process) for smart harvest by tracking the right stage, according to quality indicators, to pick the fruits. Since our offered solution is relatively cheap, we also see a potential use of this device by (3) consumers and markets that want to know the quality of the fruits they are buying/selling. Finally, since tomato has been employed as a model fruit for the scientific community, particularly in the studies on fruit development, (4) the scientific community dealing with fruit quality can also benefit from such a device because it allows assessing quality indicators in many fruits in a cheap and fast way while the fruits are still on the plant (can be used to study fruit quality changes throughout the ripening process, which is of great interest for many researchers studying environmental and meteorological effects on fruit quality indicators).
Kanter Ido
An underlying mechanism for successful deep learning (DL) with a limited deep architecture and dataset, namely VGG-16 on CIFAR-10, was recently presented based on a quantitative method to measure the quality of a single filter in each layer. In this method, each filter identifies small clusters of possible output labels, with additional noise selected as labels out of the clusters. This feature is progressively sharpened with the layers, resulting in an enhanced signal-to-noise ratio (SNR) and higher accuracy. In this study, the suggested universal mechanism is verified for VGG-16 and EfficientNet-B0 trained on the CIFAR-100 and ImageNet datasets with the following main results. First, the accuracy progressively increases with the layers, whereas the noise per filter typically progressively decreases. Second, for a given deep architecture, the maximal error rate increases approximately linearly with the number of output labels. Third, the average filter cluster size and the number of clusters per filter at the last convolutional layer adjacent to the output layer are almost independent of the number of dataset labels in the range [3, 1,000], while a high SNR is preserved. The presented DL mechanism suggests several techniques, such as applying filter’s cluster connections (AFCC), to improve the computational complexity and accuracy of deep architectures and furthermore pinpoints the simplification of pre-existing structures while maintaining their accuracies.
